Jev-powered email cleanup, with you in control.

What is the Jev AI model?

Jev is a model from TypeSafe AI for structured decisions. TypeSafe calls this category “System One”: an application provides context and defined questions, and the model returns typed results for the application to use. Its focus is classification and decision signals rather than writing free-form replies. Read TypeSafe’s introduction to Jev.

The official API supports bounded questions such as choosing an option, assigning a score, or estimating a yes/no probability. This makes it possible to ask separate questions about a message instead of relying on a paragraph that says whether to remove it. See the official TypeSafe quick start.

SiftSafe is an independent product powered by Jev. It is not an official TypeSafe or Google product.

Why use Jev for Gmail cleanup?

An email sender alone tells you very little about whether a message is worth keeping. A store might send a discount, a receipt, and an account security alert. Deleting everything from that store treats three different situations as one.

SiftSafe uses Jev to evaluate intent and other signals separately. A promotional classification is only one input; it does not automatically make an email a cleanup candidate. Transactional, financial, security, conversation, and future-reference signals can change the outcome.

AI classification becomes a reviewable decision

• Recommended: the application’s policy identifies a cleanup candidate for you to review.
• Needs Review: uncertain or incomplete signals need attention and do not silently become recommendations.
• Protected: a protection rule keeps the message out of selectable cleanup groups.

The result is organized by sender and intent. A store’s promotions can appear in one group while its orders remain protected. You can inspect groups and preview samples before deliberately making a selection.

What Jev confidence does—and does not—mean

Confidence is a model signal about a classification, not a promise that you will never need the email. A well-formed response can still be wrong about the message. SiftSafe validates responses, applies deterministic policy, and keeps a separate review path for uncertainty.

A timeout, invalid response, unsupported locale, or exhausted retry does not authorize cleanup. Such cases go to review or block the operation. Jev does not get to bypass protection rules, select messages on your behalf, or make a Gmail API call.

We do not publish a SiftSafe accuracy percentage or mailbox scan-speed claim here. Provider benchmarks do not establish how a particular Gmail account will be classified, and an individual’s definition of valuable mail can differ from a model’s prediction.

How much email data does the AI see?

SiftSafe starts with Gmail metadata and checks local protection rules before building a minimized classification request. Eligible requests can contain features such as sender domain, redacted subject features, age, locale, and protection signals. Complete message bodies are not sent by default.

Requests go through SiftSafe’s decision service to Jev. Gmail access tokens remain in the extension, and SiftSafe does not persist email content on its servers. Scan state and recovery metadata are kept locally in your browser. The Privacy Policy explains processing providers, retention, and your controls.